I had very similar thoughts to this piece at the beginning of the year but I took a bit of a different approach with the tool I built, Kan [0]. Rather than markdown files in directories like OP went with, I went for plain json files stored in a `.kan` directory, with the primary interface being not the files themselves but the CLI and an embedded frontend which could be run with `kan serve`. Agents work really well with the CLI and the frontend provides a nice snappy UI for humans to interact with and manage their tasks in a kanban format, all offline and acting entirely on files in the repo's VCS (somewhat like Typemill).

Overall it's been a great workflow and it works great with agents, it's the main way I manage tasks with agents, I agree the 'tickets in the repo' approach works really well.

[0] https://github.com/amterp/kan